Indian hospitals should not become an invisible cleanup crew for someone else’s AI
By Arunima Rajan
In an interview with Arunima Rajan, Tim Mobley, President, Connext Global, argues that AI agents in healthcare run on a global workforce of exception handlers, quality reviewers and escalation teams, and that Indian hospital operators should refuse to let their staff become "an invisible cleanup crew" for systems designed elsewhere. Connext provides the offshore oversight services he describes.
Connext builds and manages dedicated offshore teams in the Philippines, India and Colombia for midsized companies, including the human staff who review outputs, handle exceptions and manage escalations around AI systems in healthcare.
Before founding the company in 2014, Mobley was president of a Honolulu dental group with, he says, around 200 employees and 10,000 patient encounters a month. From 2009 he built offshore teams in India and the Philippines to handle the group's billing and eligibility work, and spent years as an offshoring customer before starting his own firm.
Take the digital health deployment you worked on. Before the AI agents went in, how many staff handled scheduling, refills and intake, and where are those people now?
Our client, an AI-first digital health startup, was preparing to launch its product and had only two core team members. It didn't have the human operating infrastructure needed to support the platform, and that is what Connext helped with: adding that infrastructure around the AI.
There was not a group of schedulers who lost their jobs when the agents arrived. The AI was already central to the business model, and the human team was added to make the technology operational, reliable and scalable.
What work exists now that did not exist before: reviewing agent output, handling escalations, correcting errors, retraining? How many full-time equivalents does that take?
Many roles are being created around AI today: handling exceptions and complex scenarios, validating agent outputs, supporting interactions that require judgement or empathy, identifying gaps in responses, flagging inconsistencies and providing structured feedback that improves the agents over time. Humans also support onboarding, training, performance monitoring and the operational processes surrounding the platform. The hidden workforce behind an AI agent includes quality assurance, escalation management, workflow design and continuous product feedback.
What proportion of patient interactions does the AI complete without a human touching it, and what proportion gets escalated, in production rather than the design target?
The operating model separates high-volume, repeatable interactions from situations involving exceptions, ambiguity, empathy or judgement. An AI agent may complete a straightforward scheduling request independently, while a refill request involving missing information, an exception or patient uncertainty should move to a trained person.
At Connext, we believe the meaningful production metric is not simply how many interactions the AI touches without a human. We analyse how many interactions are completed correctly without later correction, repeat contact or downstream rework. A high automation percentage can look impressive while hiding poor handoffs or additional work elsewhere in the process. The goal should be to automate the interactions the system can handle reliably, and make the human handoff fast, informed and accountable when it cannot.
You ran a 200-person healthcare organisation. If someone brought you this deployment, what would you have asked that these buyers did not?
My first question would be: what problem are we actually trying to solve? Is it patient wait time, staffing capacity, after-hours coverage, cost, accuracy or something else? If the goal is unclear, it becomes very easy to call an implementation successful when all the AI did was handle a large number of interactions.
Then I would ask where accountability sits. Who owns the outcome when the agent gives an incomplete answer, misses context or sends a patient down the wrong path? What requires human review and triggers an escalation? How quickly does a person take over, and can they see the full history of the interaction?
The right answers come from knowing what combination of technology and people gives us the best patient outcome with clear accountability.
Where does the escalation work sit geographically, and at what cost relative to the US staff? To be plain with my readers: how much of this is offshoring rather than automation?
Moving the same job from the US to a lower-cost country would not make the strongest model. What companies need is an around-the-sun operation in which AI and global teams keep the workflow moving across time zones.
A healthcare organisation might use a team in Colombia for real-time overlap with its US staff, while teams in India or the Philippines continue reviewing outputs, handling exceptions and preparing unresolved cases while the US team is offline. When the US workday begins again, the information is already organised and the highest-priority cases are ready for action. Connext's India model, for example, is designed to create that day-and-night continuity.
I won't deny that the cost advantage is real, though it is not as large as it used to be. Depending on the role, country, shift and compliance requirements, global teams can cost less than equivalent US staffing. Connext commonly structures teams to produce labour savings of roughly 40%.
An offshore position does not represent a US job that disappeared. The global team extends operating hours, absorbs rising volume or gives the onshore team capacity to focus on more complex, patient-facing responsibilities.
The honest label is not automation alone. It is AI supported by an around-the-sun global workforce. The technology provides scale, but trained people in different markets provide continuity, judgement and accountability.
What has gone wrong in these deployments: a failure that reached a patient, or a category of interaction you had to pull the agents out of?
The most common failure comes from lack of oversight, and it is usually not one dramatic event. It is an AI system added to an existing workflow without redesigning the process around it. The technology produces an answer that looks complete but lacks an important detail. It fails to recognise that a situation is unusual. It sends an interaction to the wrong queue, or leaves the next employee to reconstruct the context. Each mistake may look small, but at scale they create rework, slower resolutions and a loss of trust.
In healthcare, I would be especially cautious about allowing an agent to operate independently when the interaction involves clinical judgement, emotional sensitivity, ambiguous instructions or an exception outside a clearly defined process. The critical capability is recognising when the system should stop and bring in a person.
Does the human oversight requirement fall as the models improve, or does volume rise to fill the capacity? What have you actually seen?
Human oversight does not simply disappear as the model improves. It changes, but it shouldn't disappear. As AI becomes more reliable, teams may spend less time reviewing every routine output. But that is when companies give the system more volume, more workflows and more complex responsibilities, and human work moves toward exception handling, performance monitoring, compliance, root-cause analysis and redesigning the process when the business or model changes.
Buyers initially hesitated to invest in people because they expected the technology to handle the entire workflow. As they used the tools, they began to understand where humans were still required, and started asking for integrated models in which AI covers volume spikes, nights and weekends while trained teams manage escalations.
If your goal is zero oversight, you will fail. The goal should be higher-value oversight. Better AI should allow people to move from checking every transaction to improving the entire system.
My readers run Indian hospitals, the receiving end of this work rather than the buying end. What should an Indian hospital operator understand about where the oversight layer ends up, and what skills it demands?
Indian hospital operators should not allow their teams to become an invisible cleanup crew for AI systems designed somewhere else.
The oversight layer should be treated as a strategic capability, with access to the underlying workflows, performance data and product teams. The people reviewing AI need enough authority to stop an interaction, correct a process and identify recurring failures. That is why we emphasise that clearing the queue is the wrong metric.
The skills are becoming more sophisticated. These teams need healthcare operations knowledge, critical thinking, quality assurance, data literacy, compliance awareness, root-cause analysis and the judgement to determine when automation has reached its limit. They also need to understand how prompts, workflows and escalation rules affect the final outcome. AI adoption is creating roles such as AI trainers, model evaluators, governance specialists, automation professionals and systems integrators.
India is well positioned for this work. But the opportunity is not merely to receive lower-value tasks from Western hospitals. India should build the AI operations and governance teams that design, manage and continuously improve the system. The oversight layer should be measured and managed as part of the hospital's core operation, regardless of where it sits geographically.
Got a story that Healthcare Executive should dig into? Shoot it over to arunima.rajan@hosmac.com—no PR fluff, just solid leads.